Macaron-V1 is an open agent-model family designed for experiential intelligence, enabling systems to learn from real-world experiences and continue improving after deployment. The architecture centers on two goals: adaptation through recursive self-improvement of model-harness pairs, and collaboration via a Mixture-of-LoRA (MoL) system that selects specialist adapters per user turn.

  • Flagship Macaron-V1-Venti combines a 744B GLM-5.2 base with four LoRAs for chat, agent, coding, and GenUI tasks.
  • Macaron-V1-Tall uses a Qwen3.6-based 50B model for local deployment using the same MoL design.
  • The system employs Model-Harness Co-design and the MindForge agentic RL framework for recursive self-improvement.
  • Infrastructure includes the MinT post-training platform, LongStraw for long-context RL, and stability techniques for sparse MoE models.

The report validates Macaron-V1 on Personal Intelligence, GenUI, and general capability benchmarks against frontier baselines, though compounding gains from continual learning remain open questions.